Research map / Graph learning / Knowledge and temporal graphs

Mixed Geometry Message and Trainable Convolutional Attention Network for Knowledge Graph Completion

AAAI 2024 · Graph learning › Knowledge and temporal graphs · approach: Inductive knowledge graph completion · 19 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MGTCA, a knowledge graph completion model combining a mixed hyperbolic/hypersphere/Euclidean message function, a trainable convolutional attention network that interpolates GCN and GAT behavior, and a mixed-geometry scoring function, evaluated on FB15k-237, YAGO3-10 and WN18RR.

Problem

Knowledge graph completion with GNNs suffers from data dependence (GCN vs GAT performance depends on local neighborhood structure, requiring costly pre-validation) and message limitation (Euclidean-only message functions fail to capture KG structure).

Approach

Generates neighbor messages jointly in hyperbolic, hypersphere and Euclidean spaces; uses trainable coefficients alpha and beta so that one attention layer can smoothly switch between GCN, GAT and convolved-attention behavior; and scores triples with a prediction function plus a similarity function summed over the three geometries.

Key results

• MGTCA reaches MRR 0.393 and Hits@10 0.583 on FB15k-237, improving over the prior best (GIE/MRGAT-level, MRR 0.373 to 0.393 per the text).
• On YAGO3-10 MGTCA reports MRR 0.586 and Hits@10 0.721; on WN18RR MRR 0.511 and Hits@10 0.593.
• Ablations show removing the mixed geometry message function, TCAN, or mixed geometry scoring function each lowers performance on all three datasets.

Measured effects reported in the paper

Surprising findings

• Learned alpha/beta coefficients show the first and last layers behave like GCNs while the third layer approximates KGCAT, suggesting layer-specific GNN types.

Limitations

• Gains on FB15k-237 are reported against a baseline table where some numbers are copied from original papers rather than rerun.
• Ablation of MGMF replaces it with MRGAT's message function rather than a full alternative, so the attribution is partly indirect.

Experimental setup

Datasets: FB15k-237 YAGO3-10 WN18RR

Benchmarks: FB15k-237 YAGO3-10 WN18RR

Metrics: MRR Hits@1 Hits@3 Hits@10

Baselines: TransE ConvE RotatE CompGCN HittER LTE-ConvE RotatE-IAS MRGAT GreenKGC CompoundE MuRP MuRS

Builds on: R-GCN CompGCN MRGAT GAT MuRP

Closest papers by meaning

• End-to-End Structure-Aware Convolutional Networks for Knowledge Base Completion (AAAI 2019 · similarity 0.88)
• Mixed-Curvature Multi-Modal Knowledge Graph Completion (AAAI 2025 · similarity 0.86)
• Knowledge Graphs Can be Learned with Just Intersection Features (ICML 2024 · similarity 0.85)
• Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion (AAAI 2025 · similarity 0.85)
• Geometry Awakening: Cross-Geometry Learning Exhibits Superiority over Individual Structures (NeurIPS 2024 · similarity 0.85)
• KAA: Kolmogorov-Arnold Attention for Enhancing Attentive Graph Neural Networks (ICLR 2025 · similarity 0.84)